Environmental test data prediction method for small sample deletion

Abstract

The invention discloses an environmental test data prediction method for small sample missing, which comprises the following steps of: 1, acquiring environmental test original data of a test target to form an original data matrix; performing abnormal value judgment on the original data by adopting a two-step judgment method to generate a first processing data matrix; according to the Euclidean distance between the parallel samples, determining missing data in the first processing data matrix by adopting a nearest neighbor method, and performing filling to generate a second processing data matrix; according to the second processing data matrix, based on a paired double-sample mean T test method, determining whether the performance of the test target has significant change; and when it is judged that the performance of the test target has significant change, training an environmental test data prediction model by using the second processing data and adopting a bagging ensemble learning method, and performing data prediction. The method plays an important supporting role in small sample missing environment test data filling, performance parameter prediction, normal use guarantee of products and materials thereof and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of environmental testing, and in particular relates to a method for predicting environmental testing data for small sample missing. Background Art

[0002] Environmental testing is an activity conducted to maintain reliable functions and performance in all environments of expected use, transportation or storage. Generally, the product or its main materials are exposed to natural or artificial environmental conditions to withstand environmental effects, to examine the impact of the environment on its functions or performance, and to evaluate the environmental adaptability of the product and its materials.

[0003] Different products have different requirements for performance and functions, and require different types of environmental tests, such as natural environment atmospheric exposure test, environment-operating condition coupling test, environment-fatigue / constant stress coordinated test, marine atmosphere-vibration acceleration simulation test, black box exposure test, glass frame test, etc., which result in different measurement data under different test requirements, forming a complex environmental test data system, and there are a large number of small sample data in environmental test data. At the same time, the environmental test cycle includes several days, months, and years. In the long span, there are improper human operation, abnormal collection instruments or detection machines, etc., which lead to various situations such as errors, deviations, mistakes, and missing environmental test data, which are difficult to be directly used for mining and analysis.

[0004] The current prediction methods for environmental test data with small sample missing are mainly based on the corrosion and aging mechanisms of different materials, and the analysis and research of specific materials and specific environments. However, such methods are usually not scalable and require the classification and sorting of small sample data first, and then combined with mechanism research. Therefore, we make full use of the characteristics of environmental tests, use artificial intelligence algorithms such as machine learning, and combine data characteristics to construct a universal prediction method for environmental test data with small sample missing, which is used to accurately predict the environmental test performance of products and their materials, and is of great significance to supporting environmental adaptability evaluation. Summary of the invention

[0005] The purpose of the present invention is to provide a method for predicting environmental test data with small sample sizes that is missing. The method can effectively fill in the missing data for small sample environmental tests by establishing universal method steps, and perform data prediction based on statistical analysis and machine learning methods to obtain environmental test performance parameters of products and their materials, thereby providing accurate input for environmental adaptability evaluation.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A method for predicting environmental test data with small sample missing, comprising:

[0008] Step 1: Obtain the original data of the environmental test of the test target to form the original data matrix X. If there is missing data, it is replaced by the set characters, where:

[0009]

[0010] Indicates the number of replicates at test time t n Performance parameters at time t, i∈[1,M], n ∈[t1,t N ], M represents the total number of parallel samples, N represents the total number of sampling times during the test, t1 and t N Represent the first sampling time and the Nth sampling time respectively;

[0011] Step 2: Use the two-step judgment method to judge the outliers of the original data and generate the first processed data matrix;

[0012] Step 3: According to the Euclidean distance between parallel samples, the nearest neighbor method is used to determine the missing data in the first processing data matrix and fill it in to generate the second processing data matrix;

[0013] Step 4: Based on the second processed data matrix, determine whether the performance of the test target has changed significantly based on the paired two-sample mean T test method;

[0014] Step 5: When it is judged that the performance of the test target has changed significantly, the second processed data matrix is ​​used to train the environmental test data prediction model using the bagging ensemble learning method, and the environmental test data prediction model is used to perform data prediction.

[0015] Furthermore, in step 2, a two-step judgment method is used to judge the outliers of the original data and generate a first processed data matrix, including:

[0016] According to the environmental test performance parameters of the test target, the detection threshold is determined. According to the detection threshold, a preliminary judgment is made on whether there are abnormal values ​​in the performance parameters in the original data, and the abnormal values ​​are marked;

[0017] According to the Grubbs test method, the original data after preliminary judgment is tested to determine whether there are outliers and mark them;

[0018] The data marked as abnormal values ​​are replaced with set characters, thereby generating first processed data.

[0019] Furthermore, in step 3, according to the Euclidean distance between the parallel samples, the nearest neighbor method is used to determine the missing data in the first processed data matrix and fill it in, so as to generate a second processed data matrix, including:

[0020] S31: Calculate the Euclidean distance between the parallel sample containing missing data and all remaining parallel samples, that is,

[0021] make

[0022] Calculate the Euclidean distance between two parallel samples, that is, calculate X i and X j The Euclidean distance between represents the performance parameter vector of the ith parallel sample at all test times, Represents the performance parameter vector of the jth parallel sample at all test times;

[0023] S32: According to the Euclidean distance, the K nearest neighbor parallel samples of the parallel sample containing the missing data are determined, and the missing data are filled with the mean of the data corresponding to the test time of the missing data in the K nearest neighbor parallel samples, that is, if the kth parallel sample is at t n If there is missing data at time, then the K nearest neighbor parallel samples of the kth parallel sample are used to calculate the value of t n The missing data is filled with the mean of the data corresponding to the time. After all the missing data are filled, the second processed data matrix is ​​formed.

[0024] Further, in step 4, according to the second processed data matrix, based on the paired two-sample mean T test method, determining whether the performance of the test target has changed significantly includes:

[0025] S41: Acquire the data of all parallel samples corresponding to time t1 and the data of all parallel samples corresponding to time t N The corresponding data of all parallel samples generate double samples S A and S B ,Right now:

[0026]

[0027] S42: Calculate double sample S A and S B The difference ΔS is:

[0028]

[0029] S43: Calculate T value, where:

[0030]

[0031] represents the average value of all data in ΔS, S ΔS represents the standard deviation of all data in ΔS;

[0032] S43: Query the T test critical table and compare it with the T value to determine whether the performance of the test target has changed significantly.

[0033] Furthermore, in step 5, a bagging ensemble learning method is used to train an environmental test data prediction model and the environmental test data prediction model is used to perform data prediction, including:

[0034] S51: constructing an environmental test data prediction model, wherein the environmental test data prediction model includes at least two weak evaluators established in parallel, and outputting the results in a comprehensive manner in a bagging manner;

[0035] S52: Using t1 to t N-1 All parallel sample data corresponding to time are used as training sets, and t N The data of all parallel samples corresponding to the time are used as the validation set to train the environmental test data prediction model;

[0036] S53: Using the trained environmental test data prediction model, the average of the prediction results of all weak evaluators is taken as the final prediction result.

[0037] Furthermore, the at least two weak estimators include at least two of linear regression, ridge regression, LASSO regression, elastic regression, Bayesian ridge regression, support vector machine regression, and multilayer perceptron regression.

[0038] The beneficial effects of the present invention are:

[0039] The present invention adopts a two-step processing method (i.e., a method combining the Grubbs test method and threshold judgment) to eliminate outliers, determines the data matrix in combination with the characteristics of environmental test data, uses the nearest neighbor algorithm to fill in missing values, and finally uses the paired two-sample mean T test to judge significant changes and bagging ensemble learning to achieve data prediction.

[0040] This model is applied to environmental test data prediction, has good adaptability, can effectively use the relationship between parallel samples to fill in missing data, and is more scientific and has mathematical significance compared to the previous direct use of mean filling. Moreover, the model in the present invention is easy to understand and plays an important supporting role in filling in environmental test data with small sample missing, predicting performance parameters, and ensuring the normal use of products and their materials.

[0041] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0043] Figure 1 It is a schematic flow chart of a method for predicting environmental test data with small sample missing. DETAILED DESCRIPTION

[0044] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the preferred embodiments are only for illustrating the present invention, rather than for limiting the protection scope of the present invention.

[0045] Figure 1 This is a schematic flow chart of a method for predicting environmental test data with small sample loss. Figure 1 As shown, the method includes:

[0046] Step 1: Obtain the original data of the environmental test of the test target to form the original data matrix X. In the original data, if there is missing data, replace it with the set character at the corresponding position, where the set character can be None or NAN, or any other suitable character.

[0047]

[0048] Indicates the number of replicates at test time t n Performance parameters at time t, i∈[1,M], n ∈[t1,t N ], M represents the total number of parallel samples, N represents the total number of sampling times during the test, t1 and t N Represent the first sampling time and the Nth sampling time respectively;

[0049] Step 2: Use the two-step judgment method to judge the outliers of the original data and generate the first processed data matrix.

[0050] This step 2 may include:

[0051] First, the detection threshold is determined according to the environmental test performance parameters of the test target. Based on the detection threshold, a preliminary judgment is made on whether there are abnormal values ​​in the performance parameters in the original data, and the abnormal values ​​are marked.

[0052] For example, the tensile strength of 6061 aluminum alloy is generally around 240MPa-290MPa. If the data exceeds 350MPa, the data is considered abnormal and the abnormal data is marked.

[0053] It should be noted that the threshold set here can be a single threshold or a double threshold, and there is no restriction here. The selection of the threshold value is related to the specific performance parameters that need to be considered. The threshold is determined based on historical experience and professional knowledge, and can be set manually or automatically, and there is no restriction here.

[0054] Next, after the preliminary judgment, the original data after the preliminary judgment can be tested according to the Grubbs test method to determine whether there are outliers and mark the outliers;

[0055] Then, the data marked as abnormal values ​​are replaced with a set character (eg, None or NAN), thereby generating first processed data.

[0056] In some embodiments, the marked outliers may be directly removed.

[0057] For the Grubbs test method, the commonly used scheme is to arrange a set of data in ascending order. Generally, the abnormal value appears at the minimum or maximum value, and is calculated using formula (1).

[0058]

[0059] Where s is the standard deviation of the data, is the average value of the data, X0 represents the data that may be an outlier (i.e., data that may be an abnormal value), and the maximum value (for example, the maximum or minimum value in a set of data) is given priority. The G value is calculated and compared with G(n,P) in the Grubbs critical value table, where n in G(n,P) represents the sample size and P is the set confidence level. Table 1 is the Grubbs critical value table.

[0060] Table 1 Grubbs critical value table

[0061]

[0062] If G>G(n,P), it means that the value is an outlier, otherwise it is a normal value.

[0063] For small sample data, due to the small amount of data, it is not necessary to sort them in order, and the G value of all the data in the group can be directly obtained and then the critical value can be determined. In the invention, the G value can be obtained for the data of all test times corresponding to each parallel sample, and the G value corresponding to each data is compared with G(n,P) in the Grubbs critical value table. If G>G(n,P), it means that the value is an outlier, that is, an abnormal value, and the value is marked.

[0064] In some embodiments, the data after the second step of judgment can be further manually reviewed to confirm whether the data should be eliminated or retained. The eliminated data is replaced by None or NAN. Due to the small sample size of data, the data cannot be directly eliminated by numerical test anomalies.

[0065] Through the above step 2, the data in the matrix obtained are all normal values ​​or values ​​replaced by set characters (for example, None or NAN). Then, the positions marked with set characters (for example, None or NAN) need to be filled with data.

[0066] In step 3, the missing data in the first processed data matrix can be determined and filled using the nearest neighbor method according to the Euclidean distance between the parallel samples to generate a second processed data matrix.

[0067] This step 3 may include:

[0068] S31: Calculate the Euclidean distance between the parallel sample containing missing data and all remaining parallel samples, that is,

[0069] make

[0070] Calculate the Euclidean distance between two parallel samples, that is, calculate X i and X j The Euclidean distance between represents the performance parameter vector of the ith parallel sample at all test times, Represents the performance parameter vector of the jth parallel sample at all test times.

[0071] If X i and X j If there is no missing data in the i and X j The Euclidean distance between the two parallel samples is taken as the Euclidean distance between the two parallel samples;

[0072] If X i or X j If any of the vectors in X contains missing data, i and X j The data corresponding to the test time with missing data are removed. i and X j For example, when X i In t n1 There are missing data in time, that is Data is missing (replaced by None or NAN). The None or NAN at this position needs to be replaced from X i , and X j Middle and tn1 The data corresponding to the time (whether it is replaced by None or NAN or the real data ) to be eliminated; on the contrary, if in X j If there is missing data at a corresponding time point in X j Remove this data and remove X i The data at the corresponding time point in . i Perform a elimination, then X i The length of X is reduced by 1. j The length of is also reduced by 1. By replacing X i and X j After the elimination operation, the final X i and X j There are no missing data in either, and, since X i and X j The data at the same time point in the above equations are removed, so the final X i and X j (ie updated X i and X j ) also have one-to-one correspondence in time.

[0073] Next, determine the updated X i and X j The Euclidean distance d Ω , Euclidean distance d Ω The calculation method of is the conventional Euclidean distance calculation method between two vectors, which will not be described here.

[0074] Then, according to the updated X i and X j The Euclidean distance d Ω , determine the X before the update i and X j The Euclidean distance D between them is:

[0075]

[0076] Among them, N and N' respectively represent the X before and after the update. i or X j The number of data in .

[0077] S32: According to the Euclidean distance, determine the K nearest neighbor parallel samples of the parallel sample containing the missing data, and use the mean of the data corresponding to the missing data time in the K nearest neighbor parallel samples to fill the missing data, that is, if the kth parallel sample is at t n If there is missing data at time, then the K nearest neighbor parallel samples of the kth parallel sample are used to calculate the value of t nMore specifically, if the third parallel sample has missing data at t2, and after calculation, the first, second and fourth parallel samples are the nearest neighbors of the third parallel sample, then the data of the first, second and fourth parallel samples at t2 (i.e., X 1t1 , X 2t1 and X 4t1 ) to fill in the missing data of the third parallel sample at time t2.

[0078] Among them, the choice of K value can be based on To determine, M represents the total number of parallel samples, and K can take the smallest possible value. If the total number of parallel samples is 2, that is, M = 2, then K can only be 1, that is, only the data of a parallel sample can be used to fill in the missing data in another parallel sample.

[0079] Only when the performance of materials / products undergoes significant changes after environmental testing can they be used for model construction, analysis and prediction. The initial performance and the last test performance data are taken as double samples, and a mean T test is performed to determine whether there is a significant change. Specifically, in step 4, based on the second processed data matrix, based on the paired double sample mean T test method, determine whether the performance of the test target has changed significantly. This step 4 may include:

[0080] S41: Obtain the data of all parallel samples corresponding to time t1 (i.e., the data of the first test parallel samples) and the data of all parallel samples corresponding to time t2. N The corresponding data of all parallel samples (the last test parallel sample data) generate double sample S A and S B ,Right now:

[0081]

[0082] S42: Calculate double sample S A and S B The difference ΔS is:

[0083]

[0084] S43: Calculate T value, where:

[0085]

[0086] represents the average value of all data in ΔS, S ΔS represents the standard deviation of all data in ΔS;

[0087] S43: Query the T - test critical table and compare it with the T - value to determine whether there is a significant change in the performance of the test target. Table 2 is the two - sided T - test critical value table.

[0088] Table 2 Two - sided t - distribution critical value table

[0089]

[0090]

[0091] Assume H0: ( and are the means of S A and S B respectively), there is no difference in the paired - sample means, that is, there is no significant change.

[0092] Compare the T - value with T(n, P) (the specific data is obtained by querying the T - test critical value table, see Table 2). If T < T(n, P), then accept the hypothesis H0, that is, there is no significant change in the paired - samples; otherwise, reject the hypothesis H0, that is, there is a significant change in the paired - samples.

[0093] Next, in step 5, when it is determined that there is a significant change in the performance of the test target, then use the second - processed data, adopt the bagging ensemble learning method to train the environmental test data prediction model, and use the environmental test data prediction model for data prediction.

[0094] This step 5 may include:

[0095] S51: Construct an environmental test data prediction model. Among them, the environmental test data prediction model includes at least two weak estimators established in parallel and comprehensively outputs the results in a bagging manner. Among them, the at least two weak estimators may include at least two of linear regression, ridge regression, LASSO regression, elastic regression, Bayesian ridge regression, support vector machine regression, and multi - layer perceptron regression (with 1 hidden layer).

[0096] S52: Use all the data of the parallel samples corresponding to the time from t1 to t N-1 as the training set, and use all the data of the parallel samples corresponding to the time of t N as the validation set to train the environmental test data prediction model;

[0097] S53: Use the mean of the prediction results of all weak estimators in the trained environmental test data prediction model as the final prediction result, that is:

[0098]

[0099] Among them, represents the final prediction output of the environmental test data prediction model, which is the average prediction value of multiple weak evaluators; n model represents the number of weak evaluators; represents the predicted output of the i-th weak evaluator.

[0100] It should be noted that if the performance degradation law is more complicated, reduce the use of the five types of weak evaluators, such as linear regression, ridge regression, LASSO regression, elastic regression, and Bayesian ridge regression, and use other complex algorithms if necessary. The final prediction model can predict data at multiple future time points, and there is only one prediction value at one time point, that is, using multiple parallel sample data and prediction models to predict the overall trend or parallel sample mean.

[0101] In some embodiments, after the model is trained, the performance of the model can also be analyzed. For example, the error between the measured data and the predicted data can be quantitatively analyzed through indicators such as mean square error MSE, correlation coefficient r2, and average error E (as shown in the following formula 5) to evaluate the prediction performance of the model.

[0102]

[0103] Where N represents the number of performance tests (i.e. the total number of test times, i.e., from t1 to t N The number of sampling times included in represents the average value of the measured data during the i-th detection, It represents the predicted value predicted by the model during the i-th detection. Due to the use of bagging ensemble learning, the i-th detection data is the mean of the outputs of multiple weak evaluator models, that is, the mean obtained by formula (4).

[0104] The following is a detailed description through specific embodiments.

[0105] Step 1: Collect and organize the laboratory environment acceleration data of a certain type of rubber compression deformation rate, as shown in Table 3. The missing positions in Table 3 indicate that the corresponding data are missing.

[0106] Table 3 Laboratory environment accelerated data of compression deformation rate of a certain type of rubber

[0107]

[0108]

[0109] The data are original data without any deletion. It is recommended to collect environmental test data with more than 3 parallel samples. Such data are valuable for filling missing values. If there are only 2 parallel samples, this method is completely inapplicable. If there are only 3 parallel samples, this method can be used but the effect is similar to directly taking the mean to fill missing data.

[0110] The data matrix is ​​formed based on the original data as follows, and the missing data is replaced by NAN:

[0111]

[0112] Since the data dimensions are consistent, no standardization or normalization is required.

[0113] Step 2: Remove outliers.

[0114] First of all, all data are consistent with the actual situation of rubber. The compression deformation rate usually decreases over time and ranges from 0-100, which is in line with professional knowledge and does not require any processing.

[0115] Next, the test data of different parallel samples at each test time are taken as a group. The first group of data is [78.79,78.21,76.40,76.76,77.37,78.53,76.33,78.72,79.61,77.51,79.03,82.90,83.01,82.08,81.01], and the second group of data is [71.73,71.84,70.18,70.15,73.06,73.82,71.75,72.34,74.25,75.50,73.42,74.43,77.39,77.98,79.20,79.06], and so on.

[0116] The results of the Grubbs test for the first set of data are shown in Table 4. The G values ​​of all data are less than G (15, 99%), and there are no significant anomalies, so all are retained. Since there is no marked abnormal data, there is no need to manually review the marked abnormal data.

[0117] Table 4 Results of Grubbs test for the first set of data

[0118]

[0119]

[0120] It should be noted that the 15 in G(15,99%) in the first row of Table 4 indicates that the sample size (i.e. the total amount of data after removing missing data) is 15, that is, the total number of non-missing data in the first group of data is 15; the short dash “-” in Table 4 indicates missing data and the G value is not calculated due to missing data, so the data cannot be retained.

[0121] Step 3: Fill missing data based on the nearest neighbor method.

[0122] First, fix the data matrix. The row vector of the original data matrix is ​​the compression deformation rate data of different parallel samples under the same test time. It needs to be transposed. The fixed data matrix is:

[0123]

[0124] The row vectors X1, X2, X3... respectively include the compression deformation rate data of different parallel samples at all test times 1, 2, 3, 5, 14...

[0125] Then, the Euclidean distance between any two row vectors is calculated (it should be noted that any two row vectors here also include row vectors with missing data and other row vectors), forming a 16×16 Euclidean distance matrix, which is symmetric about the diagonal:

[0126]

[0127] Select the K value, the number of parallel samples is 16, and the K value is 3.

[0128] Finally, determine the three nearest neighbors of the missing items and fill in the missing items. For example, if there are missing values ​​in row vector X2, it is closest to row vectors X1, X3, and X4, so the missing value of the third day of parallel sample 2 is filled with the mean of the third day data of parallel samples 1, 3, and 4, that is, x = (68.87 + 66.50 + 67.45) / 3 = 67.61.

[0129] The complete data matrix is ​​as follows:

[0130]

[0131] Step 4: Analysis of significant changes based on paired two-sample mean T test.

[0132] Take the first test parallel sample data and the last test parallel sample data as double sample S A and S B :

[0133]

[0134] Calculate the difference ΔS of the paired two samples:

[0135]

[0136] Calculate the T value.

[0137]

[0138] By looking up Table 2, we can get T>T(16,0.05)=2.120, which means that the sample has changed significantly. In this example, the data changes are obvious, but in some cases the changes are not significant and need to be judged by this method.

[0139] Step 5: Bagging ensemble learning realizes environmental test data prediction.

[0140] The data from the 1st, 2nd, 3rd, 5th…84th days are taken as the training set, and the data from the 112th day is taken as the validation set. Weak evaluators such as linear regression, ridge regression, LASSO regression, elastic regression, Bayesian ridge regression, support vector machine regression, and multi-layer perceptron regression (hidden layer is 1) are established in parallel. The weak evaluators predict the data from the 112th day. Table 5 shows the predictions of the data for each test time using the method of the present invention, wherein the data for the 112th day is used to evaluate the model results, and the predictions for the previous test times are used to evaluate the training process. As shown in Table 5, the results are comprehensively output in a bagging manner, and the average prediction of each weak evaluator is taken. As the final prediction result, the measured value mean refers to the mean of multiple parallel sample data at each test time, which serves as the benchmark data.

[0141] Table 5. Ensemble learning prediction results

[0142]

[0143] Next, the model prediction performance is evaluated.

[0144] Calculate the predicted mean The error with the mean of the measured value is MSE=13.13 and the correlation coefficient is r 2 =0.90, average error E=3.13, which realizes the quantitative evaluation of the model prediction performance.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for predicting environmental test data with small sample loss, characterized in that: include: Step 1: Obtain the original data of the environmental test of the test target to form the original data matrix X. If there is missing data, it is replaced by the set characters, where: X itn Indicates the number of replicates at test time t n Performance parameters at time t, i∈[1,M], n ∈[t1,t N ], M represents the total number of parallel samples, N represents the total number of sampling times during the test, t1 and t N Represent the first sampling time and the Nth sampling time respectively; Step 2: Use the two-step judgment method to judge the outliers of the original data and generate the first processed data matrix; Step 3: According to the Euclidean distance between parallel samples, the nearest neighbor method is used to determine the missing data in the first processing data matrix and fill it in to generate the second processing data matrix; Step 4: Based on the second processed data matrix, determine whether the performance of the test target has changed significantly based on the paired two-sample mean T test method; Step 5: When it is judged that the performance of the test target has changed significantly, the second processed data matrix is ​​used to train the environmental test data prediction model using the bagging ensemble learning method, and the environmental test data prediction model is used to perform data prediction.

2. A method for predicting environmental test data with small sample loss according to claim 1, characterized in that: In step 2, a two-step judgment method is used to judge the outliers of the original data and generate a first processed data matrix, including: According to the environmental test performance parameters of the test target, the detection threshold is determined. According to the detection threshold, a preliminary judgment is made on whether there are abnormal values ​​in the performance parameters in the original data, and the abnormal values ​​are marked; According to the Grubbs test method, the original data after preliminary judgment is tested to determine whether there are outliers and mark them; The data marked as abnormal values ​​are replaced with a set character, thereby generating first processed data.

3. A method for predicting environmental test data with small sample loss according to claim 1, characterized in that: In step 3, according to the Euclidean distance between the parallel samples, the nearest neighbor method is used to determine the missing data in the first processing data matrix and fill it in, so as to generate a second processing data matrix, including: S31: Calculate the Euclidean distance between the parallel sample containing missing data and all remaining parallel samples, that is, Calculate the Euclidean distance between two parallel samples, that is, calculate X i and X j The Euclidean distance between represents the performance parameter vector of the ith parallel sample at all test times, Represents the performance parameter vector of the jth parallel sample at all test times; S32: According to the Euclidean distance, the K nearest neighbor parallel samples of the parallel sample containing the missing data are determined, and the missing data are filled with the mean of the data corresponding to the test time of the missing data in the K nearest neighbor parallel samples, that is, if the i-th parallel sample is at t n If there is missing data at time, then the K nearest neighbor parallel samples of the ith parallel sample are used to calculate the value of t n The missing data is filled with the mean of the data corresponding to the time. After all the missing data are filled, the second processed data matrix is ​​formed.

4. A method for predicting environmental test data with small sample loss according to claim 3, characterized in that: The determination of the Euclidean distance between two parallel samples involves: If X i or X j contains missing data, then X i and X j The data corresponding to the test time with missing data are removed. i and X j Make updates; Determine the updated X i and X j The Euclidean distance d Ω , According to the updated X i and X j The Euclidean distance d Ω , determine the X before the update i and X j The Euclidean distance D between them is: Where N' represents the updated X i or X j The number of data in .

5. The method for predicting environmental test data with small sample loss according to claim 1, characterized in that: In step 4, based on the second processed data matrix, based on the paired two-sample mean T test method, it is determined whether the performance of the test target has changed significantly, including: S41: Acquire the data of all parallel samples corresponding to time t1 and the data of all parallel samples corresponding to time t N The corresponding data of all parallel samples generate double samples S A and S B ,Right now: S42: Calculate double sample S A and S B The difference ΔS is: S43: Calculate T value, where: represents the average value of all data in ΔS, S ΔS represents the standard deviation of all data in ΔS; S43: Query the T test critical table and compare it with the T value to determine whether the performance of the test target has changed significantly.

6. A method for predicting environmental test data with small sample loss according to claim 1, characterized in that: In step 5, the bagging ensemble learning method is used to train the environmental test data prediction model and the environmental test data prediction model is used to perform data prediction, including: S51: constructing an environmental test data prediction model, wherein the environmental test data prediction model includes at least two weak evaluators established in parallel, and outputting the results in a comprehensive manner in a bagging manner; S52: Using t1 to t N-1 All parallel sample data corresponding to time are used as training sets, and t N The data of all parallel samples corresponding to the time are used as the validation set to train the environmental test data prediction model; S53: Using the trained environmental test data prediction model, the average of the prediction results of all weak evaluators is taken as the final prediction result.

7. A method for predicting environmental test data with small sample loss according to claim 6, characterized in that: The at least two weak estimators include at least two of linear regression, ridge regression, LASSO regression, elastic regression, Bayesian ridge regression, support vector machine regression, and multilayer perceptron regression.

Images